Methods and apparatus for reducing leakage in distributed deep learning

2. The method of claim 1, wherein reducing the distance correlation reduces invertibility of the raw data from the activation outputs.

3. The method of claim 1, wherein the loss function is a sum of a weighted distance correlation and a weighted categorical cross entropy.

6. The method of claim 1, wherein the loss function is α1DCOR(X,Z)+α2CCE(Ytrue, Y), where DCOR is distance correlation, CCE is categorical cross entropy, X is the raw data, Z is activation outputs of the intermediate layer, Ytrue is true labels for the raw data, Y is predicted labels for the raw data, α1 and α2 are scalar weights, and n is the number of samples of the raw data.

8. The method of claim 7, wherein reducing the distance correlation reduces invertibility of the one or more specific features of raw data from the activation outputs.

9. The method of claim 7, wherein the training is on all of the raw data.

10. The method of claim 7, wherein the loss function is a sum of a weighted distance correlation and a weighted categorical cross entropy.

13. The method of claim 7, wherein the loss function is α1DCOR(X,Z)+α2CCE(Ytrue, Y), where DCOR is distance correlation, CCE is categorical cross entropy, X is one or more features but not all features of the raw data, Z is activation outputs of the intermediate layer, Ytrue is true labels for the raw data, Y is predicted labels for the raw data, α1 and α2 are scalar weights, and n is the number of samples of the raw data.

16. The method of claim 14, wherein the method further comprises initializing, with weights and biases learned during the training, weights and biases of layers that are in a distributed neural network and are calculated by the client computer.